Pederzolli et al. VOL. 11, NO. 3/MARCH 2019/J. OPT. COMMUN. NETW.

15

Path-Based Fragmentation Metric

and RSA Algorithms for Elastic
Optical Networks
Federico Pederzolli, Domenico Siracusa, Andrea Zanardi, Gabriele Galimberti,
Domenico La Fauci, and Giovanni Martinelli

Abstract—Flexi-grid technology has emerged as the guard-bands) [2], thereby increasing the total potential
evolution of fixed-grid DWDM core optical networks, ena-
network throughput.
bling more efficient utilization of spectral resources and
leading to higher overall network throughput. However, The flexi-grid paradigm, however, also exacerbates the
it suffers from an aggravated form of the spectrum frag- problem of spectrum fragmentation [3], especially when
mentation problem that affects fixed-grid networks. This
work presents a novel path-based metric to better evaluate connections may be deactivated (i.e., in dynamic traffic
the fragmentation of spectral resources in flexi-grid optical scenarios). In fixed-grid optical networks, fragmentation
networks, which considers that not all links of a path may happens only when a particular channel (λ) is available
contain enough free frequency slots and the likelihood that only on some, but not all, links of a path. Flexi-grid net-
an end-to-end spectral void is large enough to house a con-
nection. Furthermore, two families of heuristic routing
works also suffer from fragmentation, but the problem in
and spectrum allocation (RSA) algorithms that preven- this context is more complex because the available slot
tively attempt to minimize the value of an input fragmen- blocks (i.e., uninterrupted ranges of free 12.5 GHz fre-
tation metric are presented. These are used to evaluate quency slots) may be too small to support an incoming con-
the effectiveness of multiple metrics, both introduced in
this work and from the existing literature, via simulations.
nection, either (i) directly on one of the links or (ii) due to
We find that both proposed families can perform better, them overlapping only partially on the links of a path, thus
in terms of blocking probability and achievable network creating a smaller end-to-end usable slot block, none of
throughput, than the well-known K-shortest paths routing which could happen in plain DWDM. These situations arise
with a first fit spectrum allocation policy, with one of them
whenever a new connection reduces the size of a relatively
trading performance for lower complexity, and that the
proposed fragmentation metric outperforms the existing large free slot block on a link to a small one or cuts it into
ones. two smaller ones. In a meshed network, this then affects all
Index Terms—Flexi-grid optical networks; Spectrum the paths crossing that link in addition to the path of that
fragmentation; Routing and spectrum allocation. connection.
In this work, we present a novel path-based metric
for measuring spectrum fragmentation in flexi-grid optical
I. INTRODUCTION networks employing super-channels or multiple modula-
tion formats, which we call the “wasted-unusable-free ratio”
(WUFR). It is based on accounting for both (i) frequency
I n order to cope with perennially increasing capacity
demands, large transport optical networks are evolving
from fixed-grid DWDM technology, where each optical
slots that are only free on some, but not all, of the links
of a path and (ii) slots that are free on all links of a path
lightpath is assigned to a fixed-size channel (λ) in a 50 GHz but belong to an available slot block that is too small with
spaced, ITU-defined grid, to elastic flexi-grid technology, respect to the number of slots needed to carry the super-
where each connection can be assigned a variable-width channels supported by the network (on which we assume
channel encompassing a number of adjacent units on a to have some advance knowledge, as explained in more de-
much tighter 12.5 GHz spaced grid (a spacing that could tail in Section III). We compare the effectiveness of this
be further shrunk as technology evolves) [1]. This new metric against two other metrics from the literature by
paradigm allows placing co-routed optical signals at the incorporating them into two routing and spectrum assign-
Nyquist condition, creating so-called “super-channels” with ment heuristic algorithms based on exploring the whole
improved spectral efficiency (due to avoiding intermediate solution space of new assignments, which we show is a fea-
sible task for centralized control planes with abundant
computational resources (e.g., [4]), although it could even
Manuscript received September 18, 2018; revised November 15, 2018;
accepted December 4, 2018; published January 28, 2019 (Doc. ID 346266). be used in fully distributed control planes (for example,
F. Pederzolli (e-mail: [email protected]), D. Siracusa, and A. Zanardi are based on the GMPLS suite of protocols [5]).
with FBK CREATE-NET, Trento, Italy.
G. Galimberti, D. La Fauci, and G. Martinelli are with Cisco Photonics
A note on terminology: in accordance with previous sci-
Srl, Vimercate, Italy. entific publications, throughout this document, we will use
https://doi.org/10.1364/JOCN.11.000015 the term “slot” to refer to the (currently 12.5 GHz) units of

1943-0620/19/030015-11 Journal © 2019 Optical Society of America

16 J. OPT. COMMUN. NETW./VOL. 11, NO. 3/MARCH 2019 Pederzolli et al.

the flexible spectral grid rather than the IETF terminology, which computes the ratio between the number of possible
which uses the same word to name a variable-width demands that can be placed on the available spectral voids
channel. in a certain link and the number that could be placed if all
This paper is organized as follows: Section II gives an those free slots were contiguous, as encoded in Eq. (3)
overview of previous works that deal with the problem of (quoted from [11]):
spectrum fragmentation. Then, Section III presents our
novel metric to measure fragmentation in flexi-grid optical P P
i k f i DIVGk
networks and gives a concrete example of how it and two ABP  1 − P , (3)
j all_free_slots DIVGj
other metrics from the literature are computed. Section IV
describes two simple heuristic routing and spectrum
allocation (RSA) algorithms based on minimizing an in- where f i is the number of spectral slots in void i, Gx is the
put fragmentation metric as well as a short analysis of set of possible demand sizes, and DIV is the integer division
their complexity and real-world applicability. Afterwards, operator. This metric is similar to FR, differing only in the
Section V describes the simulation environment we utilized way the capacity of spectral voids is computed (knapsack
to evaluate the proposed metric and algorithms and details for FR, number of fittable connections of each size for
the results of our simulations. Finally, Section VI concludes ABP). Yet another metric, called “fragmentation factor,”
the work by summarizing our findings. is proposed in [12], based on discouraging the use of voids
at the outer edges of the spectrum of each link but rather
II. PREVIOUS WORKS reusing whatever empty slots are available in the middle
where possible. The authors of [13] attempt to define an
analytical model for predicting fragmentation; however,
Many works in the literature address the problem of their model is limited to networks employing a single
spectrum (de)fragmentation, which can be decomposed modulation level and a random fit spectrum allocation
in two related subproblems: measuring fragmentation strategy, which severely limit its applicability. Another
and reducing it by defragmenting a set of connections.
model, based on Markov chains, is presented in [14] for
There is no consensus in the literature regarding how to two-service-size networks.
measure fragmentation in optical networks: Different
Likewise, many techniques have been proposed to
authors propose different metrics, with no clear indication
address spectrum fragmentation, which can be broadly
regarding which of these best captures the dynamics of this
classified as reactive, proactive, or preventive.
process (see [6,7] for an overview of multiple metrics). For
instance, if we denote as the spectral void a set of contigu- Reactive de-fragmentation techniques are triggered
ous free slots bordered by occupied slots on a certain spec- when an incoming connection request is blocked, resulting
trum representation, then the authors of [8] define the in an attempt to reorganize the resources assigned to active
fragmentation as the ratio of the maximal data rate provi- connections, possibly minimizing the number of affected
sionable using the available spectral voids and that provi- connections. For example, in [15] a blocked demand may
sionable if the same number of free slots was contiguous trigger the iterative shifting of some (as few as possible)
(i.e., part of a single, large void). More formally, denoting of the active connections closest to the spectral resources
as Gi the ith block of contiguous free frequency slots and that are to be assigned to blocked demand (which may,
as v
Gi  the maximal data rate provisionable using block in turn, trigger further shifts, up to some recursion depth
Gi , then the metric proposed in [8], called the “fragmenta- or limits of the available spectrum). Another example is
tion ratio” (FR), is defined as described in [16], where blocked connections may trigger
P the shifting of already established ones occupying the nec-
i v
Gi  essary spectral resources using a make-before-break ap-
FR  1 − P : (1)
v
i Gi  proach. Another approach, described in [17], and later
extended in [18] for entire super-channels, is known as
The authors of [9] use, instead, a metric based solely on the “push–pull,” and involves shifting an active connection in
number of spectral voids, while the authors of [10] present the spectrum without service interruptions. This technique
a fragmentation metric based on the well-known Shannon’s is, however, only applicable when the shifting connection is
entropy (which considers both voids’ size and number); not rerouted and when there are no other connections ac-
more formally, quoting from [10], the authors compute tive on any traversed link on the spectrum between the
their fragmentation index Hfrag as starting and final positions of the shifting connection.
Proactive defragmentation techniques are similar to re-
X
N
Di D active ones but are run either periodically or as soon as a
Hfrag  − ln i , (2)
i1
D D threshold of total network fragmentation, as measured by
some metric, is reached. An example of the former can be
where D is the total number of frequency slots in the spec- found in [19], where each optical connection periodically
trum of a link or path, and Di is the size, in slots, of the ith checks whether it can shift itself toward smaller frequen-
block of contiguous used or unused spectrum (hence, N is cies, resulting in an overall tendency of compacting used
the number of such blocks). The authors of [11] use, in- spectral resources toward the “left side” of the spectrum.
stead, a metric called access blocking probability (ABP), Another can be found in [20], where active connections
Pederzolli et al. VOL. 11, NO. 3/MARCH 2019/J. OPT. COMMUN. NETW. 17

are reorganized every time a certain number of them are advance-reservation connection demands are considered,
terminated. and propose a technique to preemptively rearrange sched-
Unlike the former two cases, preventive techniques do uled (but not yet active) connections to make room for new
not involve resource defragmentation but rather the care- advance-reservation demands. The same problem is tackled
ful placement of new connections during the computation in [32]. The authors of [33] consider instead the fragmenta-
of solutions to the RSA problem in order to minimize frag- tion deriving from the slow upgrade on transponder technol-
mentation (e.g., by attempting to exactly fill the gaps left by ogy, by accounting for the potential future fragmentation
a terminated connection) and therefore reduce or eliminate of different super-channel sizes during RSA. Another angle
the need to execute complex and potentially disrupting on spectrum fragmentation is discussed in [34], which con-
defragmentation processes later. There are many examples siders the fragmentation stemming from instantiating
of such fragmentation-aware RSA schemes in the litera- whole virtual optical networks at a time. Furthermore,
the introduction of the space dimension in space division
ture, such as [9,10], where the authors proposed a heuristic
multiplexing (SDM) optical networks can lead, for some
RSA scheme based on minimizing an entropy-based frag-
types of SDM hardware, to the problem of space fragmen-
mentation metric. Another work using an entropy-based
tation, as discussed in [35].
metric is [21], where the authors use the “utilization
entropy” of a path (the average entropy of each slot on that There is one aspect in common among all of these works:
path, computed on the sequence of links of the path) they measure fragmentation on a binary vector represen-
to measure fragmentation, and adaptively change the tation of spectral occupation, representing either a link,
minimum size of connections based on the overall frag- thus failing to recognize misalignments of spectral voids
mentation on the network, reducing it as fragmentation in- on successive links, or a path, thus capturing the end-
creases. This implies that large demands are increasingly to-end spectral voids but losing information pertaining
split and inversely multiplexed into multiple smaller ones to free slots that are only available on some of the links of
(diluting the benefits of super-channels) to increase their that path. We propose a metric, described in the following
probability of fitting into the available resources as frag- section, which considers both the quality of end-to-end
mentation increases. spectral voids and the presence of free slots on a subset
of the links of a path, and we show that it can be used
Another approach for preventing fragmentation and, in
to better prevent the effects of fragmentation (namely, con-
particular, lessening its effects for larger demands, which
nection blocking) compared with previous metrics.
are more vulnerable to it, is to separate the available spec-
tral resources into partitions, and preferably assign them
to connections of a certain size, as described by, e.g., [22]. III. MEASURING FRAGMENTATION WITH THE
A variation of this approach is described in [23], where WASTED-UNUSABLE-FREE RATIO METRIC
partitions are strictly assigned to a particular demand
size, with an extra shared partition to accommodate any
We propose the use of a novel fragmentation metric,
demand that does not otherwise fit. Similar ideas are dis-
which considers the ratio between the unusable or even
cussed in [24,25].
partially usable slots on a path and the total free slots.
The authors of [26] use, instead, an approach based on In order to properly define it, we first need to define a
always leaving spectral voids that can contain an integer few auxiliary concepts: We define an available frequency
number of new connections, on the assumption that their range on a link (AFRL ) as a range of frequencies that is
possible sizes are known in advance. Another example can not allocated to any connection in that link, i.e., a group
be found in [27], where the authors propose to use a metric of contiguous free frequency slots on a link (≥1 slot).
very similar to that proposed in [8] and to use a heuristic This is equivalent to what we earlier called a spectral void
that finds the assignment that minimizes such a metric on (also known as spectral hole or spectral gap in the litera-
all the links of the network, potentially also giving prefer- ture) but restricted to the case where the considered spec-
ences to utilizing slots that are already utilized in some trum models a single link. We also define an available
other part of the network. A similar approach is described frequency range on a path (AFRP ) as a frequency range
in [28], which proposes a heuristic technique based on min- that is not allocated to any connection in any link of that
imizing the number of spectral voids that would be “cut” in path, i.e., groups of contiguous frequency slots that de-
two by the placement of a new connection. This technique is scribe the same free frequencies in all the links of the path
also applied in [29], to a distributed GMPLS control plane and can therefore be used to support a new connection on
where the RSVP-TE protocol is extended to compute poten- that path.
tial cuts during forward signaling and then the destination Given these definitions, we define the wasted-unusable-
node selects the most appropriate spectral placement of the free ratio (WUFR) metric for a certain path as the ratio be-
new connection being established based on this informa- tween the sum of (i) the “wasted slots” factor (W), which
tion and relays it back during backward signaling and, counts slots that are free only on a strict subset of the links
in [30], applied to an SDN-based control plane. of the path (more formally: which belong to at least one
Last, the problem of fragmentation in optical networks AFRL but not an AFRP ) and (ii) the “unusable slots” (U),
is not necessarily strictly limited to just the plain which account for small AFRP s, i.e., end-to-end spectral
spectrum domain. The authors of [31] describe the voids that are only big enough to support a strict subset
problem of time fragmentation, which occurs when of the possible super-channels supported by the network,
18 J. OPT. COMMUN. NETW./VOL. 11, NO. 3/MARCH 2019 Pederzolli et al.

over the number of free slots (F) on the path (i.e., the sum requirement for this to be so, we did not explore the ram-
of all free slots on each of the links of the path), as shown ifications of evolving the underlying distribution in this
in Eq. (4): work. Let us highlight the fact that p only considers the
probability of being able to fit a single incoming demand


W  U∕F if F > 0 in an AFRP ; in this way, we only need to know (or estimate)
WUFR  : (4) the probability distribution of arrivals for each demand
0 if F  0
size and not worry about all possible combinations of future
More in detail, we describe the spectrum availability on a demand sizes. This greatly simplifies calculations and does
link using a binary vector, where 0 means that a slot is free not really reduce the effectiveness of our metric; this is be-
and 1 that it is currently being used by an optical connec- cause an AFRP ’s ability to host future demands beyond the
tion. The sum of all the binary vectors of the links of a one under consideration is subject to extreme perturba-
certain path gives an integer vector S representing the slot tions: any connection establishment (which would either
availability on that path. The values of each element s ∈ S shrink or split an AFRP ) or termination (enlarging or cre-
are, by construction, confined to the range 0, N L , where ating an AFRP ), even on intersecting paths, can affect the
N L is the number of links traversed by the path under usability of an AFRP , making long-term predictions unre-
study. Note that each s directly refers to a value contained liable at best. In the degenerate case of uniformly distrib-
in the vector, not its index (in essence, we mostly treat the uted connection sizes, p reduces to computing the fraction
vectors as sets because the only structures we care about of possible super-channel sizes that fit into a certain AFRP .
are the AFRL s and AFRP s). Given this representation, we Finally, the free slots term F, defined in Eq. (5c), is the
compute the W, U, and F terms of that path using the count of free slots on the links of the path and can be ob-
following equations: tained from S simply by summing, over the elements s ∈ S,
the number of links in the path (N L ) minus s (which, by itself,
X N L − s if 0 < s < N L counts the links in which a particular slot is occupied).
W , (5a)
s∈S 0 otherwise The special case for F  0 in Eq. (4) is a technicality to
handle the unlikely case where the available spectral re-
sources on the path are fully utilized, which would result
X
U
jhj · N L ·
1 − p
jhj, (5b) in computing 0/0. Clearly, if there are no free slots, none can
h∈H be unusable due to fragmentation, and the operator ought
to be happy with its network (and transitioning to a
X new one).
F
N L − s: (5c)
s∈S
A. Examples of Computing Fragmentation
The wasted slots term, W, is defined in Eq. (5a) as the sum
of those elements of the vector that are neither 0 nor N L, or, As an example, imagine a simple linear network with
in more intuitive terms, the count of those slots that are four nodes, A, B, C, and D and three links, A–B, B–C, and
free only on some of the links of this path and are therefore C–D, whose spectrum occupation is depicted in Fig. 1, and
unusable for serving a connection on it. Remember that, by where incoming connections can require four or seven
construction, 0 ≤ s ≤ N L . slots (corresponding to 100 and 200 GB/s super-channels
The unusable slots term, U, is defined in Eq. (5b), where using existing coherent-detection transceivers), with
the set H is the set of all AFRP s on the path; the cardinality equal probability. We can obtain the path vector for path
of one of its elements, jhj, is the size, in frequency slots, of A-B-C-D (N L  3) by simply summing the vectors SAB ,
the AFRP h. The p function encodes the probability that a SBC , and SCD , obtaining the path spectrum vector S in Fig. 1.
single incoming connection could fit in that particular Then, we can compute the count of wasted slots, as de-
AFRP . In other words, we factor as unusable a fraction of tailed in Eq. (5a), which corresponds to summing N L minus
the perfectly usable slots of an AFRP proportional to the
probability of not being able to fit an incoming connection
into it.
Note that computing p requires knowledge of the pos-
sible sizes of the incoming connection requests (i.e., the
spectral width of the associated super-channels) and at
least approximate knowledge of their relative distribution,
so that more frequent sizes can be given more weight. Also
note that both of these parameters should be at least ap-
proximately known to operators because they depend on
their own policies and available hardware and can even
be learned “on the job” by keeping track of the size and
number of incoming demands. We assumed the distribu-
tion encoded by p to be static; while there is no strong Fig. 1. Examples of link and path spectrum vectors.
Pederzolli et al. VOL. 11, NO. 3/MARCH 2019/J. OPT. COMMUN. NETW. 19

the values in the yellow boxes (neither zeroes nor threes), B. Interpreting WUFR
obtaining W  11. These are the free slots that we cannot
use on this path because they are not free on all of its links.
Observe that, by construction, the WUFR metric for any
Computing U is a little more involved: We start with
path can exhibit values in the range [0, 1], where 0 means
U  0; the first AFRP (green group) is three slots wide, but no fragmentation (i.e., no wasted or unusable free slots),
three is too small to support any connection, so p
3  0; e.g., when there are no active connections on any link of
therefore, the path, or when all slots are in use, and where 1 means
that all available (i.e., free) slots on the links of the path
U ← U  3 · 3 ·
1 − 0  0  9  9. cannot be used to establish connections, e.g., because no
two AFRL s align on different links. Consistently with many
That is, we count all the slots of the first AFRP group, on other definitions (e.g., [3,8]), WUFR is purely a fragmenta-
all the links of the path, as unusable. Continuing with the tion metric as, unlike some other metrics, it contains no
second AFRP , which has size 6, we obtain that only half of information about the number or ratio of free and occupied
the possible connections can fit in it; therefore, slots. This fits well with our proposed usage: the insertion
of a new connection in the network, where numerous poten-
U ← U  6 · 3 ·
1 − 0.5  9  9  18. tial RSA solutions, sharing the exact same state but that
of the new connection, need to be compared. As a conse-
In other words, we consider the slots in this AFRP as only quence, on a path where all but one slot are in use, in a
half usable because it has a 50% chance of being unable to network which does not admit one-slot-wide connections,
fit an incoming connection. Finally, we compute F using the resulting fragmentation would be 1 (i.e., 100%). This
Eq. (5c), obtaining F  38. Therefore, the fragmentation in- is not a bug: the path has high utilization, to the tune
dex WUFR for path A-B-C-D is of
N − 1∕N, where N is the number of frequency slots
(likely 320 or 396), but the only available slot cannot be
WUFR 
W  U∕F 
11  18∕38  0.763: useful, due to a type of fragmentation peculiar to flexi-grid
networks.
Conversely, we could compute the amount of fragmentation Also observe that the output of WUFR has a probabilistic
on path A-B-C-D according to the FR or Hfrag metrics using interpretation: in the example above there is a 76.3%
Eqs. (1) and (2) on the binary path vector S0 derived by chance that a randomly selected free slot on one of the links
setting all s ∈ Sjs > 0 to 1 (or, equivalently, by computing of that path may not be useful, due to the fragmentation of
the element-wise logic OR of the link vectors). By applying available resources, to establish an optical connection over
Eq. (1), we find that G1 , i.e., what we called the first AFRP , it, and this holds until some connection crossing at least
has a potential provisionable data rate of 0 because it can- one of the links of the path is either removed or added.
not support any connection, while G2 , which corresponds to
Because WUFR is a path-specific fragmentation metric,
the second AFRP , can support a single 100 Gb/s super-
and because optical network control planes generally
channel. If they were not disjointed, they could support ei-
attempt to place connections on something close to the
ther two 100 Gb/s super-channels or a single 200 Gb/s
shortest possible path (to minimize the effect of physical
super-channel; therefore,
layer impairments), in order to obtain a network-wide
fragmentation measure, it is possible to simply compute
FR 
0  100∕200  0.5,
the average WUFR among the shortest paths between each
possible source-destination couple (thus preserving the
which captures the fragmentation on the end-to-end spec-
probabilistic interpretation of the metric’s output). If addi-
trum but cannot, because it is based on a binary vector, cap- tional information is available pertaining to the likelihood
ture that of slots that are only available on certain links of that an optical connection will be requested between
the path (hence the lower value). two specific nodes or the specific routing policies used, this
We can also compute Hfrag by applying Eq. (2) to S0 , simple average can be straightforwardly replaced with a
where we find that weighted one to give more importance to the busiest com-
municating source-destination couples and/or ignore the
      
4 4 3 3 4 4 effect of fragmentation on unfeasible or unused paths.
H frag  − ln  ln  ln
24 24 24 24 24 24
   
6 6 7 7 IV. FRAGMENTATION-MINIMIZING RSA SCHEMES
 ln  ln  1.56,
24 24 24 24
We propose two heuristic RSA algorithms that minimize
which, in addition of being unable to capture information a fragmentation metric:
related to partially available frequency slots on the path, is
somewhat difficult to interpret, as this metric contains in- • Minimize-Fragmentation-1 (MF1), which analyzes all
formation on how many free and used slot blocks are mixed feasible slot assignments for an incoming connection over
but does not relate this to how useful the free blocks are. just the shortest usable path (i.e., the shortest among
Note that, in [10], the authors apply Hfrag directly to link the K-shortest paths containing at least one suitably
vectors, not paths, so that is what we did in Section V. large AFRP );
20 J. OPT. COMMUN. NETW./VOL. 11, NO. 3/MARCH 2019 Pederzolli et al.

• Minimize-Fragmentation-K (MFK), which analyzes all (i.e., the number of frequency slots), and N C is the number
feasible slot assignments for an incoming connection over of slots required by the incoming connection; therefore, the
all precomputed K-shortest paths. expression
N S − N C  1 represents the count of valid slot
assignments in the case where all slots on the path are
In other words, MF1 greedily stops at the first suitable available. The complexity of these algorithms is therefore
path, while MFK always checks all K-shortest paths. The O
K ·
N S − N C  1 · CF , where CF is the complexity of
high-level pseudo-code for MF1 is shown in Algorithm 1; the F function.
MFK is nearly identical, except for the fact it lacks the
The complexity of CF depends on the chosen metric: for
stopping condition expressed in lines 11–13.
WUFR and FR, it is a function of the number of shortest-
path pairs, number of links per path, number and size of
Procedure 1. Minimize-Fragmentation-1 (MF1) pseudo-
voids per link, and possible size of connections. Hfrag is sim-
code
ilar but depends on the number and size of both contigu-
Input: Graph G, Frag. metric F, list of K-shortest paths in ously free or occupied spectrum blocks and does not
G, Demand
Src, Dst, Size. depend on the possible connection sizes. Despite this poten-
1: BestMetric ← ∞ tially high complexity, in our practical testing, running a
2: BestSlot ← −1
simple (i.e., not particularly well optimized) implementa-
3: for each path in the list do
tion of MFK with a relatively complex metric (WUFR)
4: for each valid slot assignment in the path do
on a realistic scenario (see Section V) took of the order
5: G0 ← G plus Demand
Src, Dst, Size, slot
of a couple of seconds using commodity hardware (a basic
6: if F
G0  < BestMetric then
laptop) in the worst case (i.e., an empty network). This may
7: BestMetric ← F
G0 
even be low enough to be implemented on top of an optical
8: BestSlot ← current slot assignment
9: end if node, but it is definitely feasible to run even such simple
10: end for and unoptimized algorithms on a centralized PCE or
11: if BestSlot ≥ 0 then SDN controller, which are based on powerful commodity
12: return BestSlot servers.
13: end if In addition, both algorithms are amenable to further
14: end for optimizations. For example, unless the implementation
15: return BestSlot of F (i.e., the network-wide fragmentation metric) has
Output: Index of the first slot in which to place the demand unusual side effects, evaluating F
G0  can be done inde-
or −1 if no valid solution is found. pendently for every G0 ; thus, it is trivial to partition
the search space and parallelize the search on multiple
Both algorithms evaluate all valid slot assignments on CPU cores (which we did to speed up our simulations).
the first among the K-shortest paths (precomputed using Another avenue for optimization is in the implementation
the well-known Yen’s KSP algorithm), meaning they try of the F functions: We naïvely recompute the fragmenta-
all possible placements for the smallest spectrum block tion of every shortest path (or link for the H frag metric
that can contain the incoming demand in each suitably [10]) in G0 , while exploiting well-known memoization tech-
large AFRP and then evaluate longer paths. MF1 stops niques [36], it would be sufficient to recompute that of those
as soon as one path yields a valid solution, while MFK al- paths (links) intersecting the new demand added to G0 ,
ways performs the search over all K-shortest paths. In both while using previously cached values (from G) for the
algorithms, the F (“fragmentation”) function computes others. Another optimization could be to maintain, for each
some fragmentation metric defined on the whole network. path, its pre-summed path vector and a list of pointers to
In the case of ties, the algorithms operate on a “first-best- all other intersecting paths, tentatively update only those,
fit” basis. Note that, for simplicity, slots are counted from 0 and compute the metric as a delta with respect to the pre-
rather than from an anchor frequency as in [1] (a simple vious state, which would remove the dependency on the
shift, with no loss of generality). Both algorithms are con- number of shortest paths in the network (O
n2 , where n
ceptually similar to the ones discussed in [7,12] yet slightly is the number of nodes), replacing it with the number of
more general because fragmentation needs not be com- intersecting short paths (a much lower number, in practical
puted on a per-link basis. topologies).

A. Computational Complexity Analysis B. Control Plane Requirements

Note that the proposed RSA schemes, particularly MFK, Another problem related to the implementation of the
can be computationally expensive because there can be a proposed metric and algorithms in the real world is that
large number of possible assignments to analyze. This of the information necessary to perform the computation.
is especially true in lightly loaded scenarios, where the In fact, it is necessary to know
number of possible assignments is close to the upper bound
of K ·
N S − N C  1 ≈ K · N S, where K is the K parameter (A) the state of occupation of every slot in every link of the
of KSP, N S is the number of elements of the path vector network (for all examined metrics);
Pederzolli et al. VOL. 11, NO. 3/MARCH 2019/J. OPT. COMMUN. NETW. 21

(B) the weights and sizes associated with the p function of DP-QPSK modulation, which, to the best of our knowledge,
Eq. (5b) (for WUFR) or related information, such as the also has enough reach to connect any two nodes in the
admissible super-channel sizes (for FR [8]). chosen topology without needing signal regeneration. In
this work, we do not consider impairment validation for
Existing distributed control planes may or may not dis- lightpath feasibility evaluation nor signal regeneration.
tribute A and definitely do not distribute B without appro- Simulations involve a dynamic traffic scenario, where con-
priate extensions. Both can easily be made available in a nection requests are generated according to a Poisson pro-
centralized setting. Observe that the information related cess and have an exponential holding time chosen, so as to
to A is largely the same as that required to compute solu- have a target ideal input network load, expressed in terms
tions to the RSA problem; hence, existing control planes of the fraction of spectral resources available in the net-
are already able to express it; any modifications needed work that are in use. For each strategy, 40 independent
are limited to the distribution of the whole network view. runs generating 105 connection requests for several target
The information related to B may even be static (or, at ideal input network loads ranging from 5% to 60% load
least, seldom change); hence, it may be distributed as an were performed. Measures of interest are averaged within
initial configuration rather than requiring protocol exten- each run (they are not independent within a run), and the
sions. Therefore, in a distributed setting, any overhead resulting values (which are independent) averaged be-
would be limited to the distribution of all modified link vec- tween runs to compute the 95% confidence intervals
tors to all nodes (i.e., largely negligible given the low con- of each.
nection churn in large transport networks).

V. SIMULATION SETUP AND RESULTS A. Simulation Results

The performance of the proposed metric and RSA We evaluated the performance of the proposed RSA
schemes was evaluated using a custom purpose-built algorithms (MK1 and MFK) using three fragmentation
event-driven simulation tool, employing the German na- metrics: WUFR, fragmentation ratio (FR) from [8], and
tional backbone network topology, consisting of 14 nodes entropy (ENT) from [9,10], leading to six different RSA
and 46 directed links (Fig. 2, [37]). Links are modeled using strategies: three stopping at the first feasible path (WUFR-
320 12.5 GHz frequency slots (i.e., as standard C-band MF1, FR-MF1, and ENT-MF1) and three searching on all
fibers in a system supporting 80 DWDM channels). K-shortest paths (WUFR-MFK, FR-MFK, and ENT-MFK).
Connection request sizes are chosen with a uniform distri- Note that, for these results, the network-wide values for
bution from a set of {100, 200, 300, 400} Gb/s corresponding WUFR and FR were averaged over the path vectors of
to super-channels occupying {4, 7, 10, 12} frequency slots of all shortest paths, while those for entropy were averaged
12.5 GHz, assuming guard-bands only at the borders and over the binary vectors describing all network links, as
this is suggested to be the most effective solution in
[9,10]. FR and entropy were also tested by averaging, re-
spectively, their value for each link and shortest-path vec-
tors, obtaining broadly similar results, indistinguishable
from those presented as far as the ranking of the metrics
is concerned.
We also compared the performance of these algorithms
and metrics with that of the well-known (and de facto stan-
dard) first fit (FF) strategy in the same conditions, namely,
employing KSP routing, with K  3, where paths were or-
dered based on the number of hops they contained.
Intuitively, FF chooses the first (i.e., left-most) available
AFRP in the first (i.e., shortest) path with a suitably large
AFRP .
The effectiveness of these algorithms was evaluated in
terms of blocking probability (BP), defined as the ratio be-
tween rejected and total connection requests, measured
network load, defined as the ratio between in-use and total
frequency slots, and network throughput (TP), defined as
the average capacity of all active optical circuits for an in-
put network load. Note that, for all algorithms, a request is
blocked if and only if no path among the K-shortest has a
suitably large AFRP .
Figure 3 depicts the measured BP against the input net-
work load for FF and the MFK-based algorithms, namely,
Fig. 2. German national backbone network topology. WUFR-MFK, FR-MFK, and ENT-MFK. Here, FR-MFK
22 J. OPT. COMMUN. NETW./VOL. 11, NO. 3/MARCH 2019 Pederzolli et al.

1e+00 produced nearly identical results to those of simple BP and

is therefore not shown. In fact, measuring blocking using
BBR instead of BP produced the exact same algorithm
Measured Blocking Probability

1e-01 ranking.
Let us remark on the practical significance of these results:
our WUFR-MFK (and, to a much lesser extent, FR-MFK)
1e-02 reaches the 10−2 BP threshold (i.e., 1%), sometimes cited
by operators as the point at which they devote resources to
upgrading a network, at much higher load than FF (at around
1e-03 43% load, compared with about 35% for FF), suggesting the
FF
WUFR-MFK possibility to significantly delay the need for infrastructure
FR-MFK upgrades. Likewise, for operators following the policy of
ENT-MFK
1e-04 upgrading the network as soon as the first blocking occurs,
0.1 0.2 0.3 0.4 0.5 0.6
our metric can significantly delay this occurrence, as evi-
Input Network Load
denced by the large gap (20% versus 30%) between the first
recorded blocking of FF and WUFR-MFK. These favorable
Fig. 3. Blocking probability versus input load for FF,
comparisons extend to the other metrics from the literature.
WUFR-MFK, FR-MFK, and ENT-MFK.
The behavior of the various algorithms with respect to
BP is also reflected in the manner in which the measured
performs significantly better than the baseline FF, starting
load scales with the input load, as shown in Figs. 5 and 6,
to block later (at about 30% load rather than 20%), and ex-
which depict the measured versus input network load for
hibiting lower overall blocking, at least until about 40%
FF and the MFK and MF1 algorithms, respectively. Let
load, when its BP converges to that of FF. Even better is
us remind the reader that we measure load as the ratio be-
the performance exhibited by WUFR-MFK, which blocks
tween the in-use and total frequency slots and define the
significantly less, all the way to 60% load, even when plot-
(ideal) input network load as the load at which, on average,
ted in logarithmic scale. Last, ENT-MFK performs consis-
the active connections in the network occupy the defined
tently worse than FF; we will discuss the performance of
fraction of total slots. We do this by accounting for the dif-
this metric in greater detail at the end of this section.
ferent spectral footprints (i.e., sizes) of the possible con-
Figure 4 shows instead the measured BP against the nections, their relative frequency, and the lengths of the
input network load for FF and the MF1-based algorithms, shortest paths in the network weighted by the likelihood
i.e., WUFR-MF1, FR-MF1, and ENT-MF1. Using MF1 al- that a connection is requested between a pair of nodes.
gorithms, the performance differences are far smaller than Therefore, because all presented algorithms rely on KSP,
those of MFK ones; nonetheless, WUFR-MF1 does appear in the absence of blocking or detours, the resulting mea-
to offer statistically significant better performance than sured load should be identical to the input one, thus over-
the baseline, at least at medium loads. lapping the bisecting dashed line in Figs. 5 and 6. Blocking
An additional investigation on blocking performance, would result in traces lower than the bisecting line (gener-
based on using the bandwidth blocking ratio (BBR) metric, ating less load than expected), while connections installed
i.e., the weighted average of the BP of each class of service over the non-shortest paths would push the resulting traces
(100 Gb/s, 200 Gb/s, etc.) weighted against its bandwidth, above said line.

1e+00 0.6
Measured Blocking Probability

0.5
Measured Network Load

1e-01
0.4

1e-02 0.3

0.2
1e-03
FF FF
WUFR-MF1 0.1 WUFR-MFK
FR-MF1 FR-MFK
ENT-MF1 ENT-MFK
1e-04 0
0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
Input Network Load Input Network Load

Fig. 4. Blocking probability versus input load for FF, WUFR- Fig. 5. Measured load versus input load for FF, WUFR-MFK,
MF1, FR-MF1, and ENT-MF1. FR-MFK, and ENT-MFK.
Pederzolli et al. VOL. 11, NO. 3/MARCH 2019/J. OPT. COMMUN. NETW. 23

Measured Network Throughput relative to FF

0.6 1.035
FF
1.03 WUFR-MFK
0.5 FR-MFK
1.025
Measured Network Load

ENT-MFK

0.4 1.02
1.015
0.3 1.01
1.005
0.2
1
FF
0.1 WUFR-MF1 0.995
FR-MF1
ENT-MF1 0.99
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.985
0.1 0.2 0.3 0.4 0.5 0.6
Input Network Load
Input Network Load

Fig. 6. Measured load versus input load for FF, M-WUFR-MF1,

Fig. 7. Measured TP (relative to FF) versus input load for FF,
FR-MF1, and ENT-MF1.
WUFR-MFK, FR-MFK, and ENT-MFK.

As can be expected, while the input load is low, blocking

Measured Network Throughput relative to FF

1.035
is nearly or completely absent, and the measured load FF
scales linearly with the input load. More specifically, it al- 1.03 WUFR-MF1
FR-MF1
most exactly overlaps (only slightly higher than expected, 1.025 ENT-MF1
due to occasional detours) the bisecting line for FF and all 1.02
MF1 algorithms (Fig. 6), as these all deploy connections on
1.015
the shortest possible path barring high resource exhaus-
tion. Because MFK algorithms sometimes shun shortest 1.01
paths in favor of longer ones, resulting in lower overall 1.005
blocking, this naturally entails that the load they place
1
on spectral resources is higher than their MF1 counter-
parts. This is evident in Fig. 5, where all MFK algorithms, 0.995
and especially FR-MFK and WUFR-MFK, exhibit a sig- 0.99
nificantly higher than baseline measured load. While 0.985
ENT-MFK does not appear to offer significant gains corre- 0.1 0.2 0.3 0.4 0.5 0.6
sponding to the slight increase in resource consumption, Input Network Load
FR-MFK and especially WUFR-MFK amply justify their
resource intensiveness with their increased rate of accep- Fig. 8. Measured TP (relative to FF) versus input load for FF,
tance of connections. At higher loads, where blocking starts WUFR-MF1, FR-MF1, and ENT-MF1.
to become significant, i.e., ≥10−2 (around the 35%–40%
mark, consistently with Figs. 3 and 4), the measured load the 10.45% achieved by FF. The behavior of the MF1 algo-
growth slows significantly for all algorithms, as fewer and rithms is broadly similar, as for earlier results, with about
fewer connection demands are accepted. half the gain for WUFR-MF1 and limited to no advantage
The next set of results, depicted in Figs. 7 and 8, using the other metrics.
describes the measured throughput (TP), relative to that In order to check that our results were not tied to the
of FF, against the input load of the network, again grouped topology used, a similar analysis, not depicted due to space
in terms of MFK and MF1 algorithms. Consistent with the constraints, was also performed for the Spanish National
earlier results, at low loads there is no difference between Topology [38], which, while being of similar size (geographi-
the various algorithms because blocking is too infrequent cally speaking), is much more meshed than the German
to have a significant impact. As the input load increases, one, containing 30 nodes and 112 directed links. With
WUFR-MFK exhibits significantly higher throughput than respect to the results presented in this section, this change
FF, managing to install about 3% more capacity than FF had basically no effect, with the algorithms performing
(which, for reference, achieves a total network throughput in the same manner relative to each other. The only sig-
of about 200 Tb/s at 60% load), while the other MFK algo- nificant difference was running time: As discussed in
rithms actually show a slight decrease in total throughput Section IV, a larger topology graph implies longer running
(although FR does achieve better throughput than FF at times; experimentally, a two-core laptop could compute
middling loads). This can be attributed to higher BP at high WUFR on an empty (i.e., worst case) German topology in
loads; further, while not clearly visible in Fig. 3 due to the about 7 s and about 50 s on the much more complex
log scale, the BP of FR-MFK and ENT-MFK at 60% load is Spanish one (although some of the optimizations hinted
actually higher, at 10.97% and 11.58%, respectively, than at earlier could greatly reduce this difference).
24 J. OPT. COMMUN. NETW./VOL. 11, NO. 3/MARCH 2019 Pederzolli et al.

B. Results Discussion unusable free slots on a network path. In addition, two fam-
ilies of heuristic RSA algorithms, minimize-fragmentation-
K (MFK) and minimize-fragmentation-K (MF1), were
With the evidence presented, we can draw some conclu-
presented and their computational complexity discussed.
sions about the performance of the proposed algorithms.
First, selecting a MF1 versus an MFK algorithm mainly The performance of the proposed algorithms was evalu-
seems to boil down to a trade-off between computational ated using WUFR and two metrics borrowed from the
complexity and amount of fragmentation reduction, as available literature on the subject (fragmentation ratio
the ability to utilize additional (non-shortest) paths does and entropy), against the well-known K-shortest path with
not appear to have an impact on the ranking of the metrics first fit algorithm, by means of simulations. The results
but only on the magnitude of their effects. show that WUFR is a more effective metric at capturing
By using only the shortest path with a suitably large the nuances of fragmentation than the other metrics from
AFRP , like FF does, MF1 algorithms do not meaningfully the literature and that both families of algorithms can
increase the amount of spectral resources used by the net- perform better (MFK potentially much better) than FF
work compared with FF, at least not until lower blocking in terms of both blocking probability and network through-
justifies this increase with improved throughput. This is put with some fragmentation metrics. Furthermore,
the case for WUFR-MF1, which uses 2.6% more spectral results show that MFK algorithms use consistently more
resources compared with FF, compensated by a 1.5% in- spectral resources than FF and MF1 ones because, by ex-
crease in average network throughput. ploring alternative paths, they tend to delay loading the
busiest links in the network.
In contrast, the MFK family of algorithms exhibits sim-
ilar yet more dramatic results: It can offer larger improve-
REFERENCES
ments in terms of blocking and throughput (up to about 3%
more instantiated capacity for WUFR-MFK), which are
strongly dependent on the efficacy of the fragmentation [1] “Spectral grids for WDM applications: DWDM frequency
metric being used but at the cost of using significantly more grid,” ITU-T Recommendation G.694.1, Feb. 2012.
spectral resources (almost 14% more for WUFR-MFK). [2] O. Gerstel, M. Jinno, A. Lord, and S. Yoo, “Elastic optical
This increased usage of spectral resources is the conse- networking: a new dawn for the optical layer?” IEEE
quence of two factors: 1) at higher loads, the increased net- Commun. Mag., vol. 50, no. 2, pp. s12–s20, Feb. 2012.
work throughput exhibited by some of these algorithms [3] W. Shi, Z. Zhu, M. Zhang, and N. Ansari, “On the effect
implies that more and/or larger (and more efficient) connec- of bandwidth fragmentation on blocking probability in elastic
tions are being established, and 2) at all loads, these algo- optical networks,” IEEE Trans. Commun., vol. 61, pp. 2970–
2978, July 2013.
rithms have the ability to route requests on paths that
are not strictly shortest, even if there are enough resources [4] F. Paolucci, F. Cugini, A. Giorgetti, N. Sambo, and P. Castoldi,
“A survey on the path computation element (PCE) architec-
on shorter ones, when doing so reduces the overall frag-
ture,” IEEE Commun. Surv. Tutorials, vol. 15, pp. 1819–
mentation they perceive. This is done whenever taking a
1841, 2013.
detour moves enough fragmentation from the busiest links
[5] E. Salvadori, Y. Ye, C. Saradhi, A. Zanardi, H. Woesner, M.
(i.e., the links traversed by the most paths) to the least-
Carcagni, G. Galimberti, G. Martinelli, A. Tanzi, and D. La
used ones because fragmentation on any link is weighted Fauci, “Distributed optical control plane architectures for
proportionally to the numbers of paths traversing it, so handling transmission impairments in transparent optical
that the net amount of fragmentation in the network would networks,” J. Lightwave Technol., vol. 27, no. 13, pp. 2224–
be lower than if the detour was not taken. In other words, 2239, July 2009.
MFK algorithms include a sort of implicit load balancing [6] B. C. Chatterjee, S. Ba, and E. Oki, “Fragmentation problems
mechanism performed jointly with fragmentation minimi- and management approaches in elastic optical networks: a
zation, which can have a large impact in terms of BP at survey,” IEEE Commun. Surv. Tutorials, vol. 20, pp. 183–210,
low loads. 2018.
Finally, the ENT metric appears to work rather poorly in [7] M. Quagliotti, D. C. Arango, M. Schiano, A. Carena, M.
conjunction with the proposed RSA heuristics. We note Cantono, and V. Curri, “Spectrum fragmentation metrics
and their use in optical channel allocation algorithms,” in
that, in [9,10], the metric was applied to a network filling
19th Italian Nat. Conf. on Photonic Technologies (Fotonica),
(i.e., without connection removals) scenario, which greatly
June 2017, pp. 1–4.
inflates its measurements compared with a situation where
[8] R. Wang and B. Mukherjee, “Spectrum management in hetero-
only sporadic voids are left in the “left-side” of the spectrum.
geneous bandwidth networks,” in Global Communications
Conf. (GLOBECOM), Dec. 2012, pp. 2907–2911.
VI. CONCLUSIONS [9] P. Wright, M. Parker, and A. Lord, “Minimum- and maximum-
entropy routing and spectrum assignment for flexgrid elastic
optical networking [Invited],” J. Opt. Commun. Netw., vol. 7,
In this work, we briefly described the problem of no. 1, pp. A66–A72, Jan. 2015.
fragmentation in flexi-grid optical networks and various [10] P. Wright, M. Parker, and A. Lord, “Simulation results of
definitions given for it in the scientific literature and pre- Shannon entropy based flexgrid routing and spectrum assign-
sented a new fragmentation metric, wasted-unusable-free ment on a real network topology,” in 39th European Conf. on
ratio (WUFR), which is able to account for both wasted and Optical Communication (ECOC), Sept. 2013, pp. 1–3.
Pederzolli et al. VOL. 11, NO. 3/MARCH 2019/J. OPT. COMMUN. NETW. 25

[11] D. Amar, E. Le Rouzic, N. Brochier, E. Bonetto, and C. Lepers, partition for elastic optical networks,” IEEE Photon. J., vol. 9,
“Traffic forecast impact on spectrum fragmentation in 7906413, Oct. 2017.
gridless optical networks,” in European Conf. on Optical [26] D. Borquez, A. Beghelli, and A. Leiva, “Deadlock-avoiding vs.
Communication (ECOC), Sept. 2014, pp. 1–3. greedy spectrum allocation algorithms in dynamic flexible op-
[12] R. Dutta, S. S. Krishna, R. Gowrishankar, G. J. Subhanu, tical networks,” in Int. Conf. on Transparent Optical Networks
S. S. S. Sanagapati, P. P. Behere, and S. K. Bhyri, (ICTON), July 2015, pp. 1–4.
“Fragmentation-aware routing, wavelength and spectrum [27] M. Zhang, W. Lu, Z. Zhu, Y. Yin, and S. Yoo, “Planning and
assignment (RWSA) scheme in flex-grid optical net- provisioning of elastic o-OFDM networks with fragmenta-
works,” in IEEE Int. Conf. on Advanced Networks and tion-aware routing and spectrum assignment (RSA) algo-
Telecommunications Systems (ANTS), Dec. 2017, pp. 1–6. rithms,” in Asia Communications and Photonics Conf.
[13] S. Yan, J. Kim, X. Wang, and A. Fumagalli, “Spectrum (ACP), Nov. 2012, pp. 1–3.
fragmentation analysis in a two-rate elastic optical link,”
[28] Y. Yin, M. Zhang, Z. Zhu, and S. Yoo, “Fragmentation-aware
in Advanced Photonics, 2015, paper NeM2F.4.
routing, modulation and spectrum assignment algorithms in
[14] J. Kim, S. Yan, A. Fumagalli, E. Oki, and N. Yamanaka, elastic optical networks,” in Optical Fiber Communication
“An analytical model of spectrum fragmentation in a two- Conf. (OFC), Mar. 2013, pp. 1–3.
service elastic optical link,” in Global Communications Conf.
[29] T. Fukuda, L. Liu, K.-I. Baba, S. Shimojo, and S. Yoo,
(GLOBECOM), Dec. 2015, pp. 1–6.
“Fragmentation-aware spectrum assignment for elastic opti-
[15] R. Wang and B. Mukherjee, “Provisioning in elastic optical cal networks with fully-distributed GMPLS,” in Optical Fiber
networks with non-disruptive defragmentation,” J. Communications Conf. and Exhibition (OFC), Mar. 2015,
Lightwave Technol., vol. 31, no. 15, pp. 2491–2500, Aug. 2013. pp. 1–3.
[16] Z.-S. Shen, H. Hasegawa, and K.-i. Sato, “Integrity [30] H.-C. Le, L. Liu, and S. B. Yoo, “Distributed control plane with
enhancement of flexible/semi-flexible grid networks that spectral fragmentation-aware RMSA and flexible reservation
minimizes disruption in spectrum defragmentation and
for dynamic multidomain software-defined elastic optical
bitrate-dependent blocking,” J. Opt. Commun. Netw., vol. 7,
networks,” in Optical Fiber Communications Conf. (OFC),
pp. 235–247, Apr. 2015.
Mar. 2015, pp. 1–3.
[17] F. Cugini, F. Paolucci, G. Meloni, G. Berrettini, M. Secondini,
[31] N. Wang, J. Jue, X. Wang, Q. Zhang, H. Cankaya, and M.
F. Fresi, N. Sambo, L. Poti, and P. Castoldi, “Push-pull defrag-
Sekiya, “Holding-time-aware scheduling for immediate and
mentation without traffic disruption in flexible grid optical
advance reservation in elastic optical networks,” in Int.
networks,” J. Lightwave Technol., vol. 31, no. 1, pp. 125–133,
Conf. on Communications (ICC), June 2015, pp. 5180–5185.
Jan. 2013.
[32] R. Zhu, J. P. Jue, A. Yousefpour, Y. Zhao, H. Yang, J. Zhang, X.
[18] N. Sambo, F. Paolucci, G. Meloni, F. Fresi, L. Poti, and P.
Yu, and N. Wang, “Multi-path fragmentation-aware advance
Castoldi, “Control of frequency conversion and defragmenta-
reservation provisioning in elastic optical networks,” in
tion for super-channels [Invited],” J. Opt. Commun. Netw.,
vol. 7, no. 1, pp. A126–A134, Jan. 2015. Global Communications Conf. (GLOBECOM), Dec. 2016,
pp. 1–6.
[19] D. Siracusa, A. Broglio, A. Zanardi, E. Salvadori, G.
Galimberti, and D. La Fauci, “Hitless network re-optimization [33] H. Hasegawa, T. Yasuda, Y. Mori, and K. I. Sato,
to reduce spectrum fragmentation in distributed GMPLS “Fragmentation-minimized transponder upgrading employ-
flexible optical networks,” in European Conf. on Optical ing channel bandwidth aligned slot allocation in flexible grid
Communication (ECOC), Sept. 2013, pp. 1–3. optical networks,” in Optical Fiber Communications Conf.
(OFC), Mar. 2018, pp. 1–3.
[20] M. Zhang, W. Shi, L. Gong, W. Lu, and Z. Zhu, “Bandwidth
defragmentation in dynamic elastic optical networks [34] H. Wang, X. Xin, J. Zhang, Y. Sun, and Y. Ji, “Dynamic virtual
with minimum traffic disruptions,” in Int. Conf. on optical network mapping based on switching capability and
Communications (ICC), June 2013, pp. 3894–3898. spectrum fragmentation in elastic optical networks,” in
[21] L. Altarawneh and S. Taebi, “Bandwidth granularity adapta- 21st OptoElectronics and Communications Conf. (OECC),
tion for multipath provisioning in elastic optical OFDM-based July 2016, pp. 1–3.
networks,” in Int. Conf. on Electro/Information Technology [35] R. Proietti, L. Liu, R. Scott, B. Guan, C. Qin, T. Su, F.
(EIT), May 2015, pp. 236–240. Giannone, and S. Yoo, “3D elastic optical networking in the
[22] J. Comellas, A. Asensio, M. Ruiz, G. Junyent, and L. Velasco, temporal, spectral, and spatial domains,” IEEE Commun.
“Using spectrum management techniques to differentiate ser- Mag., vol. 53, no. 2, pp. 79–87, Feb. 2015.
vices in elastic optical networks,” in Int. Conf. on Transparent [36] P. Norvig, “Techniques for automatic memoization with appli-
Optical Networks (ICTON), July 2015, pp. 1–4. cations to context-free parsing,” Comput. Linguist., vol. 17,
[23] Y. Hirota, Y. Hatada, T. Watanabe, and H. Tode, “Dynamic no. 1, pp. 91–98, Mar. 1991.
spectrum allocation based on connection alignment for elastic [37] F. Agraz, S. Azodolmolky, M. Angelou, J. Perello, L. Velasco,
optical networks,” in Asia-Pacific Symp. on Information and S. Spadaro, A. Francescon, C. Saradhi, Y. Pointurier, P.
Telecommunication Technologies (APSITT), Aug. 2015, Kokkinos, E. Varvarigos, M. Gunkel, and I. Tomkos,
pp. 1–3. “Experimental demonstration of centralized and distributed
[24] H. Chen, H. Yang, B. Luo, and L. Yan, “A joint fairness-aware impairment-aware control plane schemes for dynamic trans-
and fragmentation-reduction spectrum allocation scheme parent optical networks,” in Optical Fiber Communication
in elastic optical networks,” in Opto-Electronics and Conf. (OFC), Mar. 2010, pp. 1–3.
Communications Conf. (OECC) and Photonics Global Conf. [38] F. Rambach, B. Konrad, L. Dembeck, U. Gebhard, M. Gunkel,
(PGC), July 2017, pp. 1–2. M. Quagliotti, L. Serra, and V. Lopez, “A multilayer cost
[25] H. L. Liu, L. Lv, Y. Chen, and C. Wei, “Fragmentation- model for metro/core networks,” J. Opt. Commun. Netw., vol. 5,
avoiding spectrum assignment strategy based on spectrum no. 3, pp. 210–225, Mar. 2013.